No Time Like the Present: Agentic Test-Time Training for LLM Agents
What this work does
Updates an episode-specific adapter on agent, environment or summary text while downweighting spans repeatedly seen in earlier updates.
Connection to discovery
A baseline for learning from an evolving evidence stream without repeatedly reinforcing a stuck trajectory.
Evidence
ALFWorld and software-repair experiments compare no adaptation, static adaptation, unfiltered online updates and repetition controls, including drift diagnostics.
Scope and limitations
Token novelty is only a proxy for useful evidence. New but erroneous self-generated text can still be reinforced, and episode-specific resets differ from cross-task continual learning.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.